RL environment companies were never meant to sell to labs forever. It is undeniable what cost economics will be produced by the unit cost of intelligence decreasing, but that it is still confined to a mere 10% of GDP and sophisticated engineering companies, mean that all AI infrastructure companies doubtlessly have inane customer concentration. The TAM expansion by natural knowledge dissemination will be, by default, slow, without innovation in the deployment side. Its time to leave the nest and explore the wider ocean. Still, 80% of enterprises have never touched AI to a meaningful degree. The deployment surface for app layer companies remain extremely forward deployed in a cost prohibitive way - meaning that barring regulated industries where the unit cost of white collar work is extremely high - justifying FDE costs - average white collar work remains untouched. Its not that the work is not increasingly susceptible to the productivity gains first seen with coding - its just that there are material unscalable unit AI engineering time costs that are just not cost economic to be placed with Midwestern accounting firms. The notion that these companies ought to enter the mainstream use cases also condemns the RL environment companies that do not draw their data from realistic sources. Those that increasingly create “toy” financial models to parse out things that frontier models are poor on, will not create generalizable post training infrastructure of off the shelf datasets that will allow for easier implementation of an RLaaS contract. There are many useful problems that one learns to do well in being data connoisseurs. Here is a laundry list of these notions, which include: One figures out how training data is an attack vector, especially in light of premature mech interp solutions in market (looking at you Goodfire), and is able to offer offensive and defensive solutions One figures out how mixing recipes and offering specific datasets can lead to “selling smaller capabilities” to enterprises who’ve internalized some small model training as a product strategy. One figures out how to optimize separating necessary human feedback and work artifacts and the autonomous creation of training materials for AI such as to create the most efficient processes to create personalized app layer products One figures out how to “hire” and “recruit” agentic labor exceedingly well for a new era of labor economics where human labor oversees pools of hired machine labor for white collar tasks, in effect becoming a new class of “recruiters” As we speak - whether they like it or not - data companies are getting pulled to real world businesses because they are the last and only vectors of good training data. We will see breakouts in 2026 who will be producing free versions of quickbooks, seeding new firms to owner-operate in virgin services markets from the ground up, and seeking to capture the layer of the value chain where all value accrues - the increased output of services from efficient model selection at the performance/cost/latency curve.

This month, Poolside’s Laguna XS.2 is free to train on Prime Intellect Lab. First come, first serve while reserved capacity lasts.

What a weekend. Around 30 teams showed up to build on Laguna XS.2, and the bar was very, very high. Winners below 🏆 1st: Overthinking Machines Labs @emilfristed Pseudo-full-duplex with text-only models through dialogue modeling with silence tokens. huggingface.co/spaces/poolsi… 2nd: Coding Kernels by the Pool Charlie Masters, Evan O’Leary, Jessica Mak Laguna-Dense: a ~3B fully dense distillation of Laguna XS.2 for generating CUDA kernels from PyTorch. huggingface.co/EvanOLeary/la… 3rd: attnvq @alaradirik Attention-aware product vector quantization of KV caches. huggingface.co/spaces/adirik… Honorary mention: Laguna Vision Aaron Kazah @aaronkazah A SigLIP vision encoder resampler LoRA adapters, trained on 300k examples to give Laguna XS.2 a native visual input path. huggingface.co/poolside-lagu… Huge congrats to the winners, and thank you to everyone who hacked, demoed, judged, helped, and pushed Laguna XS.2 in directions we would not have found on our own! @nvidia @PrimeIntellect @adaption_ai @huggingface

good location to build

Such an interesting little guy

setup hell kills good RL ideas. so we’re giving researchers Laguna XS.2, @PrimeIntellect Lab, and a weekend in London to run the whole loop: tasks → evals → rewards → training → rollouts → adapters → inference 14 days to go. come touch the weights: luma.com/poolsidehackathon

King’s Cross is the Silicon Roundabout of AI ft.trib.al/Tp7pzqn | opinion

we need to see more and more of hacks like these. non slop, and real contributions that can be used after it, rather than dumped!

Poolside is hosting a 2-day model research hackathon in London. Join us to push an open-weight agent model as far as you can. RL and fine-tune Laguna XS.2, our latest-generation model, on Prime Intellect Lab. Dates: May 29–30 Partners: @nvidia @PrimeIntellect @huggingface Prize: NVIDIA DGX Spark Agents need better models. Better models need cracked researchers. Link below.

you asked for it. so we made it. say hi to siliconmania.tv/weekly each reference of the weekly tech recaps. explained. its finally time to understand all of them. enjoy.